Focality De Novo: Transfer, Reasoning, and Reinforcement in Endogenous Salience

This is the sequel to "Availability from Méconnaissance. " Paper I made variable-frame availability endogenous inside a fixed description language. This paper explains how such availability transfers to novel environments, how it interacts with strategic reasoning, how payoff feedback reshapes it, and where the explanation must stop. Features are modeled as predicates over an attribute space, and a culture is the stochastic-approximation limit of environment-by-environment salience revision. The correct aggregate operator is the annealed operator T-bar = ETE, the expectation of the per-environment logit map, rather than a logit of averaged distinctiveness; the paper isolates the Jensen trap that separates the two. Environmental diversity raises the symmetric spinodal by subtracting tangent covariance, and for two features it yields an exact annealed random-field Curie-Weiss law. The paper separates and settles five gaps. Transfer: cultural fixed points push forward to a held-out environment through applicability, deriving Schelling-style stranger coordination, cross-cultural mismatch, and transfer failure when locked features do not apply. Reasoning: a double-logit salience-and-action (quantal-response) equilibrium exists, sharp action precision recovers Bacharach rarity best replies, and false focal options can survive non-singleton strategic reasoning under an explicit margin condition. Reinforcement: with success-coupled salience, matching reinforcement entrenches false focality, quality-weighted reinforcement corrects only within a bounded region, and clustered experimentation is the escape route. Repertoire: equivariance and Goodman-style underdetermination fence the choice of feature basis, so de novo focality is determined only jointly with that basis; with bounded VC capacity, shared experience gives a finite-sample transfer bound, while unrestricted predicates recover the Goodman impossibility. Empirics: the paper states preregistered tests, verified simulations, and explicit regressions back to Paper I.